We have developed a new method for three-dimensional image reconstruction in PET, which preserves the simplicity of the two-dimensional methods. It was primarily developed for our constructed PET system consisting of two rotating scintillation cameras with a restricted axial fiel-of-view. In principle, however, it can also be used for multi-ring cameras with the interplane septa removed. The basic principle of our method is to organize the projection data so that 2D reconstruction can be used. The final image can be obtain using any available 2D reconstruction technique. The method consists of a combination of analytical and algebraical algorithms, including an iterative process with two-dimensional reconstruction of transaxial slices and deconvolution of a position dependent axial spread function. With this method we avoid the complex and time-consuming operation of 3D backprojection, but can still utilize all the detected events and take advantage of the three-dimensional information. The method has been evaluated with Monte Carlo simulations and phantom measurements.